Political questions put leading Chinese AI models on the defensive, a benchmark from Aleph Alpha concludes. Models from Alibaba, DeepSeek and Moonshot AI tended to echo official positions or decline to answer.
The German vendor markets itself as a supplier of sovereign AI to governments, so the comparison serves its own pitch. Its test spanned 967 hand-picked topics, including Taiwan, Xinjiang and Tiananmen. By its own scoring, only 17 to 41 percent of replies came back balanced.
China-related prompts are not the only place the lean shows. Qwen 3.6 opened on a neutral note when asked about American censorship, then closed by defending Beijing’s handling of internet governance.
A rival gets caught in the review too. Nvidia’s Nemotron Cascade 2 returned party-line answers 17 percent of the time, a figure Aleph Alpha links to a sliver of its training data, roughly 3,500 of 9.3 million examples, drawn from DeepSeek and Qwen.
Public-facing models in China are required to carry socialist core values, and past academic audits describe the same spillover, which researchers worry could shape how large audiences think.
The results land as Europe weighs which foreign value system its own models should sit on.
